Abstract
OBJECTIVE: To develop and validate model based on clinical, imaging, and Radiomics features for predicting disease severity and delayed recovery in Mycoplasma pneumoniae pneumonia (MPP). METHODS: This multicenter retrospective study enrolled 238 patients (training cohort), 60 (testing cohort), and 278 (validation cohort). Patients were classified into non-severe MPP (NSMPP) and severe MPP (SMPP) groups based on guideline, and further stratified post-treatment into recovery or delayed recovery groups. Radiomics features were extracted from chest CT using PyRadiomics, with Least Absolute Shrinkage and Selection Operator (LASSO) regression for feature selection. Three random forest-based predictive models were developed, including Clinical-Image, Radiomics, and Integrated. Predictive performance was evaluated via by the area under the receiver operating characteristic curve (AUC), calibration, and clinical utility. RESULTS: The Integrated model demonstrated superior discrimination for severity prediction (validation AUC: 0.784, 95% CI: 0.722-0.845) and delayed recovery (validation AUC: 0.865, 95% CI: 0.770-0.960), outperforming Clinical-Image (severity AUC: 0.771, 95% CI: 0.695-0.847; delayed recovery AUC: 0.807, 95% CI: 0.724-0.950) and Radiomics model (severity AUC: 0.710, 95% CI: 0.643-0.776; delayed recovery AUC: 0.837, 95% CI: 0.724-0.950). Integrated Discrimination Improvement (IDI) analysis demonstrated significant enhancements in the Integrated model compared to both the Clinical-Image and Radiomics models for predicting both disease severity and delayed recovery (all p < 0.05). Key predictors comprised D-dimer (severity OR = 1.371; delayed recovery OR = 4.061), systemic immune-inflammation index (delayed recovery OR = 6.607), and consolidation patterns (delayed recovery OR = 2.820). CONCLUSION: The Integrated model combining clinical, imaging, and Radiomics features enhances risk stratification for MPP severity and delayed recovery.